Francois Chaubard opened the robotics edition of YC Paper Club with a useful antidote to demo intoxication. AlphaGo looked like the missing algorithm. MuJoCo made locomotion learnable in simulation. ALOHA and UMI made imitation data look scalable. Diffusion policies and vision-language-action models then appeared to supply the general policy. Each wave produced the same forecast: next year, robotics is solved.
2026 may be the year of robotics demos. A dependable general-purpose robot is a different claim.
The gap is not that nothing works. Robots now perform remarkable, dexterous sequences in controlled settings. The problem is turning a successful rollout into a service that survives new objects, bad lighting, worn hardware, long tasks and thousands of operating hours. Chaubard organizes that gap into four walls.
Wall 1: the model must respect physical consequences
A video generator can make a plausible next frame without learning the transition law a controller needs. If a generated car drives into a store and the scene smoothly turns back into a highway, the pixels may look coherent while the physics are useless. Control needs an action-conditioned transition model: given state $s_t$ and action $a_t$, what distribution of states can actually follow?
This becomes much harder with contact and deformable objects. Cloth folds, paper creases, food changes state, friction switches regimes and small geometric errors compound. A simulator can produce unlimited experience only to the extent that its dynamics transfer. That is the sim-to-real gap: cheap synthetic scale is not automatically the same distribution as the world.
Wall 2: action representation is still an open design choice
Language has a convenient discrete unit: the token. Robot action does not. A policy might predict joint positions, torques, end-effector poses, keypoints, object poses, action chunks or latent trajectories. Each representation bakes in a different set of invariances and makes some skills easier while hiding others.
The representation determines how quickly a system learns and how well it transfers across bodies. A command meaningful for a seven-degree-of-freedom arm may not map cleanly onto a mobile manipulator or five-finger hand. Part 4 repeatedly returns to this question: MEM separates task and motor timescales; R&B-EnCoRe discovers which reasoning predicts action; SimToolReal makes object pose the common interface; world-action models learn a future-conditioned latent.
Wall 3: robots are sensory-poor
Humans manipulate through a dense skin. We estimate normal and tangential force, vibration, temperature, moisture and slip across a huge surface. We can identify a charger inside a bag without looking. A robot often gets cameras, proprioception and perhaps sparse fingertip or wrist force sensing.
That matters because manipulation is contact-rich and vision is frequently occluded by the robot's own hand. A visually impressive policy may be compensating for missing sensation with narrow priors. More data cannot always recover information the hardware never measured. Better tactile skins, force sensing and multimodal policy architectures are therefore part of the learning problem, not a peripheral hardware upgrade.
Wall 4: embodiment drifts after deployment
The same motor command does not produce exactly the same motion forever. Actuators wear, joints loosen, dust and corrosion change resistance, batteries age, payloads vary and calibration moves. A policy trained against yesterday's embodiment can become wrong without any change in the task.
This is especially damaging to an end-to-end VLA whose action mapping is implicit. Old demonstrations become partly stale. The operating system for a real fleet must detect drift, recalibrate, adapt and decide when data must be recollected. A one-day benchmark rarely measures this.
The thread connecting the five talks that follow
| Talk | Wall it attacks | Core move |
|---|---|---|
| MEM | long tasks, occlusion, repeated failure | dense visual short-term memory plus compressed textual long-term memory |
| R&B-EnCoRe | scarce and noisy supervision | keep only reasoning that helps predict the embodiment's actions |
| SimToolReal | teleoperation scale, action abstraction | train universal object-pose goal reaching in simulation |
| Teleoperation-first companies | deployment uncertainty | sell the service, learn the real requirements, then automate |
| World-action models | future dynamics versus latency | learn from future prediction, then compress the inference path |
A better test than “look at the demo”
Ask five questions instead. Does the task contain unseen contact or deformable dynamics? What action interface transfers to the next tool or embodiment? Which important state is invisible to the sensors? How does success change after weeks of wear? Can the economics survive human intervention, data collection and edge inference?
These questions do not make the demos less impressive. They locate the missing engineering. The strongest message in this opening is that robotics is not waiting for one final model. It is a coupled system of dynamics, sensing, embodiment, data, evaluation, operations and economics.
The takeaway
The decade of “next year” predictions came from mistaking a newly solved layer for a solved stack. Part 4 is more valuable when read as six different attacks on that stack. Progress is real; the proof is not a cinematic rollout. The proof is a robot that remembers, adapts, survives drift, improves from its data and performs valuable work at an affordable latency.
Sources: YC Paper Club video and full Root Access transcript.